Snowball GLM 5.3 RLVR SFT
This model continues open-athena/Grug-67B-A2B-Datakit-SFT-262K-2026.09.21 with supervised fine-tuning on 5,000 rendered GLM 5.3 RLVR conversations. The run used assistant-only loss, greedy packing into 32,768-token sequences, and 136 optimizer steps on 64 H100 GPUs.
The saved checkpoint is s3://marin-us-east-02a/marin/checkpoints/glm53-rlvr-snowball-sft-2026.09.23/hf/step-135. The model retains the base model's OpenMDW 1.1 license.